Attention-Enhanced Proximal Policy Optimization for Autonomous Robot Path Planning in Dynamic Environments

Autonomous robot path planning in dynamic, cluttered environments remains a significant problem in robotics. Traditional planning algorithms require exact prior knowledge and cannot effectively handle obstacle avoidance, while present-day deep learning (DL) techniques process all inputs uniformly, which is not suitable for efficient convergence. In this work, we propose AEPPO: Attention-Enhanced Proximal Policy Optimization for autonomous robots. AEPPO introduces a multi-head self-perception mechanism for observing spatial information with the help of attention heads, enabling the policy to focus more on dangerous object-related information. A reward function based on potential-field guidance and collision penalty is also easier to learn. In the MiniGrid and Gazebo simulated environments (static, corridor, dynamic-obstacle), the experimental results show that the success rate of AEPPO is 91.8% on MiniGrid with a 3.2% collision rate, and on the most difficult dynamic map, the success rate reaches 84.3%. Compared to five other baselines - A*, RRT, DQN, standard PPO, and Attention-DQN - training convergence results show that AEPPO reaches the target performance threshold approximately 27% faster than standalone standard PPO with much lower variance, demonstrating the effectiveness of attention-guided perception for safe and efficient robot navigation. © 2026 IEEE.